A Ready-to-Use Grading Tool for Facial Palsy Examiners-Automated Grading System in Facial Palsy Patients Made Easy

Abstract

Background: The grading process in facial palsy (FP) patients is crucial for time- and cost-effective therapy decision-making. The House-Brackmann scale (HBS) represents the most commonly used classification system in FP diagnostics. This study investigated the benefits of linking machine learning (ML) techniques with the HBS.

Methods: Image datasets of 51 patients seen at the Department of Plastic, Hand, and Reconstructive Surgery at the University Hospital Regensburg, Germany, between June 2020 and May 2021, were used to build the neural network. A total of nine facial poses per patient were used to automatically determine the HBS.

Results: The algorithm had an accuracy of 98%. The algorithm processed the real patient image series (i.e., nine images per patient) in 112 ms. For optimized accuracy, we found 30 training runs to be the most effective training length.

Conclusion: We have developed an easy-to-use, time- and cost-efficient algorithm that provides highly accurate automated grading of FP patient images. In combination with our application, the algorithm may facilitate the FP surgeon's clinical workflow.

Keywords: House-Brackmann scale; application; artificial intelligence; bell’s palsy; deep learning; facial palsy; facial paralysis; facial reanimation; smile restoration.

Figure 1

Overview of the study population and distribution of the House-Brackmann scale (HBS). The red bar is visualizing ten healthy individuals as a control group. Facial palsy (FP) patients with HBS scores of IV and VI accounted for the majority of cases, respectively.

Figure 2

Flow chart of the workflow of how to train the network (training with and without cross-validation) with the training data and validate with healthy patients.

Figure 3

Preliminary image preparation. Transformation of nine single pictures to a single composed picture.

Figure 4

Basic workflow steps. Each patient image was assigned a distinct House-Brackmann scale (HBS) before being added to the neural network.

Figure 5

The different components of the machine learning model. The network is subdivided into three parts (i.e., I, II, III).

Figure 6

Evaluation of the control group. The control group comprised of ten healthy individuals of whom only one was assigned a pathological House-Brackmann scale (HBS) score (i.e., HBS > I).

Figure 7

Exemplary application run. Sample output of the network utilizing the application.

Figure 8

Implementation of automated grading in the clinical workflow. Automated grading could be used in the preoperative planning phase, as well as for direct intraoperative assessment. Following (non-)surgical therapy, automated grading may allow for standardizing patient follow-up evaluation.